Hey HN! I'm building LlamaSim, a multi-LLM framework to simulate human behavior at scale - including election polling.
It uses gpt-4o to generate synthetic identities, and uses Cerebras's Llama 3.1 8b models to run groupchats and predictions on any event/question. I'd really appreciate any insight/advice on improving it as much as possible, as I'd love to continue building upon it! I've been brainstorming a bit with long-term memory storage (such as using mem0) and live news feeds to make it truly agentic, as well converting it into a graph-theory based approach as well!